Tesla Autopilot | AP Without Lane Lines | What Does It Follow?

Subscribe to see more videos like this!

In this video I talk about Autopilot and using it on a road without lane lines. I compare two different tests and two different software versions too.

Huge shout out to James Henderson for his support on Patreon. Click the link above and go over to Patreon to show your support!

V.2019.20.4.2

[WhiteBirdPlays]

Hey man, thanks for doing short videos that display the capabilities of autopilot. I appreciate it.

[briangonigal]

Autopilot definitely follows multiple cars ahead of you. One time I was on a (lined) road that I drive on autopilot almost every day, but going much slower than usual because ahead of the car ahead of me was a big flatbed truck with an oversized load going just about 30mph. The truck turned off onto a side street and then the car directly in front of me turned & followed the truck down that same street, and then my car on autopilot turned and followed them both! It actually successfully made the turn all on its own and went down this side street that it had passed 100 times before and always completely ignored.

[oreganorx]

I don't think you could get better training data than what Tesla gets from having 100k+ people playing around with autopilot. What an ingenious plan. Kudos to Tesla 😊

[KacperLubisz]

Answering your question requires a solid understanding of how the neural networks behind autopilot work (apologies for the length of this comment, I tried to keep it concise).


We have to start by considering the input data autopilot uses, which we know to be some combination of cameras, ultrasonics and radar, though for argument's sake let's focus on the front-facing camera. The question that we have to ask is, what information is there in an image of a road for predicting lanes? if the system is perfect, what information are you yourself using to decide which direction to steer? A neural network can be thought of as a statistical tool for finding/extracting/condensing this relevant information in such an image.


There is a lot of very obvious information in such an image: the lines on the road, the edges of the road, the car(s) in front. The key to understanding is that there is an awful lot of information in such an image that is a lot more subtle, for example: cars on the other side of the road (it can maybe even pick up on the fact that it is looking at the front of a car and can definitely infer that it can't drive into those cars), trees/foliage (if you see a curved row of trees next to the road then you might infer that the road will follow that curve), road signs (in particular think of the arrow signs you see at a tight curve, it may even learn to use those signs to mean that it needs to slow down).


So with this said, how can we answer the question if the car focuses on lane lines or cars? If we could answer this question easily we wouldn't be using a neural network to determine the answer for us. The best answer I can come up with is that it's some combination of both determined by other information that we can't understand (not from this single observation anyway).


Considering the two videos there are some other factors that we might take into consideration. Firstly, we can see the seam where the two lanes of asphalt meet in the middle of the road (very clearly at 0:12). Secondly, we can also see some form of dirt on the road in the middle of each lane, this shows areas of the road where car wheels don't drive. Next, the two videos are clearly taken at different times of day by the shadows. The different lighting conditions might influence the visibility of the first two points. Though I suspect not, the shadow on the road could have influenced the result too.


Great video though, I'd love to see more rigorous testing on this road in particular various combinations of other cars on the road (both lanes) around the corner it failed on.

[shanley]

Wow, this is important in reference to the accident in TX today. Elon tweeted that autopilot requires lane lines, which is clearly incorrect.

[sks]

In North Idaho with the same software you are using I was on a rural road with strips & then continued on autopilot onto an unlined 10 mile section& it laid down nice tight right lane blue lines with no issues. No cars in front of me & autopilot set to 3mph over 35 mph speed limit. it did drop down to 28 mph when passing through a state park speed limit. Flawless!

[joshh]

Tesla is doing a good job of identifying “drivable space”. I daily drive a stretch of unmarked two lane rural highway and test this each software update as well. 8.5 really did a good job at staying on the right side of the road with a pseudo lane. I’ve noticed now with 24.2 that it wants I center on the whole road - just like yours was doing. I’d call this a regression. Tesla really needs to step up its driving policy game. Generic lane centering isn’t going to cut it much longer. So TLDR; I don’t think that car had anything to do with the blue line behavior. Mine does the same thing now on unmarked roads.

[jaykay]

Good video. I don’t trust my EAP on those kinds of roads and doubt I ever will. It’s clearly confused wo lane markers and curbs. Elon said that curbs are FSD’s nemesis. Sadly, looks like it will take basically forever for AI to learn enough situation variables to obtain legislative approval for true urban and suburban FSD. Anxious, but skeptical.

[flymeedrone]

This is a test that missed on youtube. Good work. It seems the car doesn't realize that it should stay on the right the most it can. It should know what side it must drive. What happens if a car come from the other direction?

[DanFrederiksen]

For a while maybe 6-12 months ago it would rely a lot on following cars in situations where it had no understanding of the road but now it seems that reliance has disappeared almost completely which is good. It is clearly seeing the road without lines, which is not hard to do in these situations but it is different from the original premise of needing lines at all times.
The display classification language has to develop too since it still relies on road marking line thinking.

[drdrew]

I wonder if the Model X in front of you had stayed to the right you would have as well. Lots of obvious drawbacks if AP just follows the car in front of you. That scenario was a setup for Model X to get in head on collision and then you rear end him. It wasn’t a good spot for current state of AP but full autonomous has a long way to go to handle this situation.

Great video demo

[AndrewJamison]

How do you know the car in front had the autopilot engaged does the screen show they have it enabled

[alessandro.rossini]

I think AP it's following its "estimated" lane lines based on what the nerual network thinks, and it could be influenced by other cars

[terjes]

Tesla needs to make autopilot assume that traffic is alway going in both directions unless traffic signs says otherwise. How else is autopilot supposed to learn it?

[rivernet]

Over 8 months in my model 3, and 22000 miles, I’ve never seen anything that would make me think AP keys on the car ahead. I know this is a popular theory but I don’t think it stands scrutiny. In every other aspect, the software relies on physical markers... why would it follow an unreliable moving target?

[teslaus]

Both I think. Normally my car see this wide road, it will scream 😱 and ask my help.

[mariacobretti]

just watch Teslas autonomy day and you will know how the system works